Skip to content
PlagiatScanner.de
Examination procedures · problem · PRV-01

AI concern: which evidence should you preserve?

In brief: Preserve the submitted file, submission receipt and existing native records before changing anything. Then collect genuine drafts, version histories, research notes, source records, feedback and any AI-use logs or exports. For every item, record provenance, date, connection to the work and susceptibility to change. Never manufacture an earlier draft, backdate a reconstruction or delete an awkward intermediate version.

Evidence inventory with strengths and limitsReviewed 11 September 2026

Freeze what exists before attempting an explanation

An allegation or query can prompt frantic tidying. That response may remove the clearest trace of the actual workflow. Do not edit the submitted document, overwrite it with a cleaner copy or rename files in ways that obscure their original context. Preserve the upload confirmation, filename and visible submission time. If a service provides a native export or version-history download, use that documented function and note when the export was obtained.

Investigate copies, not originals. Separate records that pre-date the concern from material assembled afterwards. A later recollection can still help, but it must carry its real creation date. Label remembered details, uncertainty and facts that can be checked against independent records. A candid contemporary reconstruction is different from a document presented as though it existed during writing.

Authority asset: evidence inventory with claim limits

Worksheet for records already available
RecordWhat it may supportWhat it does not establish alonePreserve now
Submission and receiptThe file and time recorded by the submission routeHow its contents were producedFile, portal receipt and available properties
Drafts and version historyDevelopment of passages, structure or calculationsThe identity of every editorNative history, exports and local states
Research recordsSearch routes, selection and reading activityComplete understanding of every sourceLibrary export, reading list and annotations
Notes and outlinesQuestions, decisions and argument developmentImmutability without corroborationNative files, metadata and paper pages
AI log or conversation exportCaptured prompts, responses and visible contextCompleteness, authorship or final adoptionNative export plus final-text mapping
Feedback and correspondenceWhen a change was suggestedWho ultimately composed itEmail, comments and meeting notes

Map records to specific passages and decisions

A folder containing hundreds of files does not answer a focused concern. Assign short IDs and connect every relevant item to a chapter, paragraph, table, figure or code component. For example: “D-04 is the chapter-three state before supervisor feedback; F-02 contains the tracked suggestion; S-07 is the original publication consulted afterwards.” This creates an inspectable chain instead of an indiscriminate data dump.

For a disputed sentence, early notes, highlighted literature, multiple draft states and a documented revision can be more informative together than one screenshot. Preserve contradictions. A draft may contain an abandoned argument; a gap may reflect work on another device. State the observable fact first and any explanation separately. This lets a reviewer distinguish the record from your interpretation.

Technical indicators also need modest wording. A checksum can show that one exact file still matches the bytes from which a reference value was calculated. It cannot establish when the content was first written or who wrote it. Copying and syncing can alter file timestamps. Treat these fields as supporting context rather than automatic proof of authenticity.

Describe actual AI use rather than a vague category

If you used an AI system, neither hide that use nor enlarge it defensively. Identify the service, displayed model information, date or period, purpose, relevant instructions, selection, human changes and verification. Brainstorming, copy-editing and generating a substantive passage are different contributions. Explain which outputs were rejected, adopted or substantially transformed.

Do not attach a conversation export to an ordinary email without review. It may contain names, personal data, confidential research, credentials or third-party text. Keep the unchanged record in an approved location and ask which secure transfer route applies. If a redacted copy is permitted, record the reason and scope of each omission. The guide to assessing AI chat records compares export forms and their evidential limits.

Acknowledge missing history instead of replacing it

Not every genuine workflow produces continuous version history. A local file may have been overwritten, a platform may retain few states or rough paper notes may no longer exist. Name the gap. Then look for lawful, independent traces such as email attachments, automatic backups, reference-manager entries, supervisor comments, an electronic laboratory notebook, repository commits or calendar records. Do not access another person’s account or change a shared system merely to strengthen your account.

Never recreate an “old” draft and present it as original. Do not manipulate timestamps. A precise statement about missing material is safer than an artefact with false provenance: “No draft survives for period X. Version Z was attached to an email on date Y; I have reconstructed the steps between them from memory.” Keep recollection explicitly separate from contemporary records.

Prepare a proportionate and readable evidence package

Open with a cover sheet naming the work, submitted version, reason for the package and compilation date. Add a chronology, the inventory and a short mapping to each point raised. Include material that is relevant, authorised and intelligible. For large, sensitive or proprietary files, ask the named office for an approved transfer route and deadline.

  • Originals remain unchanged and separate from review copies.
  • Each item has an ID, provenance, date and connection to the work.
  • Recollections and later reconstructions are labelled plainly.
  • Technical indicators are described only within their real scope.
  • AI contributions are classified by function and actual adoption.
  • Protected information is not circulated without authority.
  • Gaps and inconsistent traces remain visible.

The examination procedures hub places this preservation work in a wider process. The chronology–evidence–response matrix helps turn records into a focused account. For the single foundation connection on this page, consult academic writing and evidence. The other links stay within procedure and evidence documentation.

Official foundations

  1. German Research Foundation: Guidelines for Safeguarding Good Research Practice – traceability, documentation and responsibility.
  2. ALLEA: The European Code of Conduct for Research Integrity – reliable records, transparency and responsible correction.
  3. German Research Foundation: Statement on generative models in science – responsibility and transparency when using generative systems.

Sources reviewed 4 September 2026. Current institutional requirements must also be checked for a specific procedure.